我让 AI 接管了个人网站:45 篇迁移、改版、订阅、收款,30 天实录
一个停更两个月、没有监控没有备份没有订阅没有收款的个人站,被 agent 接管 30 天后的完整交付记录:先审计后动手、先备份再改、三个坑(个人实名 SMTP / 平台稳定性 / DNS 改动)和三条可复制的做法。
这篇不是教程,是一份**交付记录**:我让一个 AI agent 接管了自己停更两个月的个人网站,从服务器权限到收款到账,30 天里真正做了什么、踩了什么坑。
先给结论:agent 能接管运维,但接管不了「判断」。所有需要拍板的时刻——改哪个版本、信不信这个平台的稳定性、要不要把收款换成赞赏码——都得人来定。
起点:一个停更两个月的站
站是 Astro 静态站,45 篇文章、59 个页面,跑在一台 2 核 2G 的腾讯云轻量服务器上。表面看正常,实际上一堆暗病:
- 任意不存在的路径都返回 200 + 首页(软 404),坏链全被藏起来了
- 首页 5 个类型入口链接写错,点进去还是首页
- og 分享图 37 页指向 404,社交分享全是裂图
- RSS 里 9 条链接 404、9 条标题带
[undefined] - 没有监控、没有备份、没有订阅、没有收款
站主(我)当时的真实状态:白天上班,晚上写小说,站扔在那没人管。这其实是大多数个人站的常态——不是不想更,是维护成本高。
agent 接管的四步
一、先审计,再动手
接管前做了一次全站只读审计:59 页 / 210 次请求,逐页检查状态码、title、canonical、og、结构化数据、内链。结论是 54 页至少有 1 项异常,但高度集中在 3 个根因上:软 404、链接拼错、生成链断裂。
这一步的价值不是「找出多少问题」,是「按根因排序」。修 3 个根因,54 页的问题一起消失;逐页修,修到天亮也修不完。
二、先备份,再改
动手前把站点目录、nginx 配置、证书、crontab 全打包,本地再存一份公开站快照。然后才改第一行配置。
后面证明这一步救了命:换版式那天 nginx 配置带 BOM 导致启动失败,回滚脚本 30 秒恢复。
三、能机器检查的,都写成脚本
接手后加了三个自动化:
- 每 10 分钟健康巡检(状态码 / 响应时间 / 证书剩余天数)
- 每周日自动备份,保留 4 份
- 每周一检查是否有新文章,有才发摘要邮件
这三个脚本加起来不到 100 行,但把「记得去看一眼」从脑子里搬出去了。
四、交付,而不是「做完」
改版上线不是终点。写清楚三件事才算交付:备份在哪、怎么回滚、以后谁维护。所以每个改动都留了回滚命令,内容生成器和部署脚本也留在仓库里。
三个真实的坑
1. 个人实名账号发不了 SMTP
订阅功能做完,卡在发信上。买域名、加解析、发信域名验证全过了,最后发现腾讯云对 2026 年 3 月后开通的个人实名账号关闭了 SMTP 发信。
换 QQ 邮箱 SMTP 十分钟跑通。教训:选方案前先确认账号资质,别把链路建到一半才发现走不通。
2. 支付平台的稳定性不能假设
本来准备用爱发电收款,认证做到一半平台打不开——从我的机器和广州服务器同时测,一个连接被拒、一个 Cloudflare 521。查了下,这类访问故障它历史上出现过多次。
最后改成「微信赞赏码 + 站内托管下载链接」:零平台依赖,付款后截图发邮件,人工回链接。平台可以换,交付必须在自己手里。
3. 域名解析要懂,但别乱动
为了做邮件认证加了 MX / SPF / DMARC / DKIM 四条记录。MX 指向腾讯企业邮,意味着这个域的收信路由被改了——虽然目前没开通邮箱,但以后要接企业邮得先想清楚。
DNS 改动不可怕,可怕的是改了不知道影响什么。
结果
- 11 项线上问题修复,59 个 sitemap URL 全部 200
- 版式从「杂志风」重排成个人博客:首页文章流、文章页目录、归档、搜索、收藏
- 邮件订阅跑通:确认信、退订、每周摘要
- 模板包上架(5 份 brief + 7 个 cron 脚本 + token 账本),微信赞赏码收款
- 新增 45 张文章封面,og 分享图全部可用
这套方法能复制吗
能,但有前提:
- 你得能拍板。agent 能给方案,不能替你承担「选错平台」的后果。
- 你得给边界。我全程用的就是站里写的那套:目标、边界、验收标准、汇报频率、回滚方式。
- 你得接受慢。30 天里真正的技术活两三天就能干完,剩下的时间都在等验证、等审核、等平台恢复。
如果你也想让 agent 接管自己的站,我把自己用的 brief 模板、cron 自检脚本和 token 账本整理成了一个包,在 /pro/。不买也没关系,站里 45 篇实战都是免费读的。
下一次更新我打算写「AI 写长文怎么防跑偏」——这 45 篇里踩的坑够写三篇。
English Version
This is a delivery log, not a tutorial: I handed my two-month-stale personal site over to an AI agent and ran it for 30 days. Here is what actually happened.
Short version: an agent can take over operations, but it cannot take over judgment. Every call — which version to ship, whether to trust a platform, when to switch payments — was made by a human.
Starting Point: A Site Nobody Maintained
An Astro static site: 45 posts, 59 pages, on a 2-core/2GB Tencent Cloud Lighthouse instance. It looked fine, but underneath:
- Any unknown path returned 200 + the homepage (soft 404) — every broken link was hidden
- Five homepage links pointed to the wrong URLs and silently fell back to the homepage
- 37 pages had og:image URLs returning 404 — social cards were broken
- RSS had 9 dead links and 9 titles prefixed with
[undefined] - No monitoring, no backups, no newsletter, no revenue
The Four Steps
1. Audit first
A read-only audit of all 59 pages (210 requests) checked status codes, titles, canonicals, OG tags, structured data and internal links. 54 pages had at least one issue, but they clustered around three root causes: soft 404s, wrong links, broken generation. Fixing three root causes beats fixing fifty-four symptoms.
2. Back up before touching anything
Site files, nginx config, certificates and crontab were archived first, plus a local snapshot of the public site. That decision paid off when a config file with a BOM broke nginx — rollback took 30 seconds.
3. Automate what can be checked
Three small automations: a health check every 10 minutes (status, response time, certificate expiry), a weekly backup with 4 copies retained, and a weekly digest that only sends when there is actually a new post. Under 100 lines total, but they moved “remember to check” out of my head.
4. Deliver, not just ship
Shipping a redesign is not delivery. Delivery means: where the backup is, how to roll back, and who maintains it later. Every change shipped with a rollback command.
Three Real Problems
1. Personal accounts cannot send email
The newsletter was built, but sending was blocked: Tencent Cloud disabled SMTP for individually verified accounts activated after March 2026. Switching to QQ Mail SMTP took ten minutes. Check account eligibility before building the pipeline.
2. Do not assume platform stability
I planned to use Afdian for payments; halfway through verification the platform went down — connection refused from one network, Cloudflare 521 from another. The plan changed to a WeChat reward QR code plus a self-hosted download link. The payment channel can change; delivery must stay in your own hands.
3. DNS changes have consequences
Email verification required MX, SPF, DMARC and DKIM records. Pointing MX at Tencent's mail servers rerouted inbound mail for the whole domain. DNS is not scary; changing it without knowing what it affects is.
Results
- 11 production issues fixed; all 59 sitemap URLs return 200
- Redesigned from “magazine” to a personal blog: article stream, TOC, archive, search, favorites
- Newsletter working end to end: double opt-in, unsubscribe, weekly digest
- A template pack (5 brief templates + 7 cron scripts + a token ledger) with QR-code payment
- 45 generated article covers; social cards finally have images
Can This Be Copied?
Yes, with three conditions: you must be the one who decides; you must give the agent clear boundaries (goal, scope, acceptance criteria, reporting, rollback); and you must accept slowness — the technical work took two or three days, the rest was waiting for verification, review and platform recovery.
Read the 45 free essays at manyou-zhe.com, or grab the template pack at /pro/.
觉得这篇有用?
订阅怪招本,每周 1–2 篇:实战手记、token 偏方、决策矩阵。不定期附模板。
